Abstract:
In the way to facilitate the scope of Computer Aided Diagnosis (CAD) into the treatment of breast cancer, which is a leading issue of concerns for women worldwide in recent times, the task of breast lesion segmentation is a very critical processing step that needs to be auto- mated. Although Digital Mammography (DM) is the most popular screening tool in breast cancer detection, Ultrasound (US) imaging has recently emerged as a popular alternative due to its non-invasive nature, real time and low cost imaging.
Breast lesion segmentation from US images using deep learning techniques is quite challenging. US images contain many fuzzy contours and false edges along with the original mask. Again, there has been shortage of publicly available large annotated datasets of Breast US images for training the deep learning model. Moreover, the introduction of adversarial training for segmentation task has been quite nascent which poses major challenges of convergence and stability issues.
We have implemented a Conditional Generative Adversarial Network (CGAN) based approach for the task of breast lesion segmentation from US Images. Specifically, the network has been designed as an upgradation to the architecture associated with CGAN by imposing multi- tasking learning in the training process. Convergence as well as stability of the newly designed model has been largely improved compared with CGAN. Also, overall performance of the segmentation task has been assessed in terms of the state of the art model such as U-Net, Pix2Pix, SegNet- cGAN. In addition to this, performance improvement has been attained for different scenarios such as different dataset, different model etc.
